Dear Adrian,
thanks for taking the time. There’s a lot to wrap my head around, your insight to the data is amazing and very accurate.
We’re currently striving to publish this, and even decided to put the focus of the paper on the hierarchical modeling. I intend to submit supplementary code and full data. I pruned the data for now, but once we’re under review, I’m happy to provide a full notebook with all detail.
Some background: what we’re doing is locomotor analysis of animals, tracking points and doing kind of “motion capturing”. The example measure I sent was “speed” (measured at the nose) while the animal is locomoting in a fast gait. This is typically a type of study with limited data and especially few subjects, because although our experiments are non-invasive and conducted under very natural conditions, this is highly regulated by authorities (which is good). We cannot get more individuals at the moment. We could analyze more runs (data points per subject), but decided against it for another reason (exclude sample size bias in comparison with commonly applied frequentist’s statistics); the twenty runs were chosen at random. The “condition” is an alteration of the extent of a plexiglas enclosure.
Now because the animal we look at is a locomotor specialist (which is usually the case), it is hard to put in a prior. Of course we have expectations, but you never know if any effect will surpass intra- and interindividual variability. Also, it is quite typical that there are individual outliers, which I hoped to model appropriately: I guess it is reasonable to expect long tails, but I have no a priori knowledge that may exclude the Normal assumption. I would definitely want to avoid loading the model with any “belief”.
Do you have any suggestions on how to argue for reasonable, more specific prior assumptions in such cases?
Thanks again for now. I’ll read up on everything you wrote and get back with questions 
Cheers,
Falk